by Hugo Bowne-Anderson (Vanishing Gradients)
I recently taught a lightning lesson with Doug Turnbull on building effective agent harnesses. Here are the things I shared with the people who came along.
Quick framing. An agent is just an LLM calling tools in a loop. A harness is everything around it: tools, context, hooks, sandboxes, memory, the loop itself.
The LLM is the brain and the harness is the body:
How To Build a General Purpose AI Agent in 131 lines of Python A coding agent and a search agent built with the same 4-step pattern. Give an LLM read, write, edit, and bash and you have a general-purpose computer-using agent. Code: hugobowne/building-with-ai.
You can also read it on O’Reilly Radar!
Building Agents That Build Themselves (with Ivan Leo, Google DeepMind, ex-Manus). Rebuilds Pi and OpenClaw ideas in pure Python: agent writes and hot-reloads its own tools, markdown memory compaction, hooks, Telegram via FastAPI, Modal sandboxing. Code: hugobowne/build-your-own-ai-assistant.
Livestream of the workshop The two-hour live build.
Build Your Own Deep Research Agent with Ivan Leo (again!) From raw Gemini API call to clarifying questions, a plan, subagents running parallel Exa searches, and a cited report. Code: hugobowne/build-your-own-deep-research-agent.
Harness Engineering: Why Agent Context Isn't Enough (with Jeff Huber, Chroma). Why large context windows aren't a silver bullet, why code-writing unlocks emergent reasoning, and the inner vs outer loop of context engineering.
AI Agent Harness, 3 Principles for Context Engineering, and the Bitter Lesson Revisited (with Lance Martin, Anthropic). Manus re-architected 5 times. Anthropic rips out Claude Code's harness as models improve. Plus the playbook: Reduce, Offload, Isolate.
15 Privacy Questions Every AI Builder Is Asking (with Katharine Jarmul). Jump to Q4. Your harness can be exfiltrated. Anything in your system prompt should be safe to publish on your website.
How To Build Your First Agentic Search Application (with Doug Turnbull). Why most of the traditional search stack falls away once the agent has the user's full narrative. A dumb keyword tool plus an agentic loop beats the baseline by 15 to 30% NDCG on Wayfair WANDS and Amazon ESCI, no training.
Episode 68: A Builder's Guide to Agentic Search & Retrieval (with Doug Turnbull and John Berryman). Hand-rolling agentic search loops, John's 5-level maturity model, why LLM-as-judge isn't the shortcut you think.